🤖 AI

Hyperdimensional Decoding of Spiking Neural Networks

This paper introduces a novel SNN-HDC decoding method that outperforms existing approaches by achieving higher accuracy, lower latency, and significantly reduced energy consumption while also demonstrating the unique capability to identify unknown classes.

Original authors: Cedrick Kinavuidi, Luca Peres, Oliver Rhodes

Published 2026-04-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cedrick Kinavuidi, Luca Peres, Oliver Rhodes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize different hand gestures, like "thumbs up," "peace sign," or "waving hello." The robot uses a special kind of brain called a Spiking Neural Network (SNN).

Think of an SNN not as a standard computer processor, but as a biological nervous system. Instead of constantly humming with electricity like a lightbulb (which is how traditional AI works), this robot brain only "fires" tiny electrical sparks called spikes when something interesting happens. It's like a room full of people who only shout out when they see something specific. This makes the robot incredibly energy-efficient, but it's also very hard to figure out what the robot is thinking just by listening to the shouts.

The Problem: How to Listen to the Robot

In the past, scientists had two main ways to interpret these shouts (spikes):

The "Counting" Method (Rate Decoding): You wait until the robot finishes looking at the gesture, then you count how many times each person shouted. The person who shouted the most wins.
- The Downside: It's slow. You have to wait for the whole performance to finish. Also, everyone has to shout a lot to get a clear count, which wastes energy.
The "First to Shout" Method (Latency Decoding): You listen for the very first person to shout. The first one wins.
- The Downside: It's fast, but it's unreliable. If the robot gets a little bit of static noise, the wrong person might shout first, and the robot makes a mistake.

The New Idea: The "Holographic Flashcard"

This paper introduces a clever new way to listen to the robot, combining it with a concept called Hyperdimensional Computing (HDC).

Imagine instead of asking "Who shouted the most?" or "Who shouted first?", you ask the robot to build a Holographic Flashcard.

The Old Way: The robot has 11 people (one for each gesture). If it sees a "thumbs up," only the "thumbs up" person shouts.
The New Way (SNN-HDC): The robot has 1,024 people (dimensions). When it sees a "thumbs up," a specific pattern of people shouts. Maybe person #1, #45, and #900 shout, while the rest stay silent. This pattern is the flashcard for "thumbs up."

Because there are so many people (dimensions), the pattern is incredibly robust. Even if a few people shout by mistake (noise) or stay silent (missing data), the overall pattern is still recognizable. It's like recognizing a friend's face even if they are wearing a hat or it's raining; you don't need every single feature to be perfect to know who it is.

Why This is a Game-Changer

1. It's Faster and Smarter (Low Latency)
With the old "Counting" method, you had to wait for the whole gesture to finish. With this new "Flashcard" method, the robot starts building the pattern the moment it sees the movement. As soon as the pattern is clear enough to match a known flashcard, it says, "I know this!" It doesn't need to wait for the end of the video. It's like recognizing a song after hearing just the first few notes, rather than waiting for the whole album to play.

2. It Saves Massive Energy
Because the robot only needs to fire a few specific sparks to build the pattern, it uses much less electricity. The paper found that this new method uses 2 to 3 times less energy than the old methods. Imagine your phone battery lasting three times longer just by changing how the brain processes information!

3. It Can Spot the Unknown
This is the coolest part. If you train the robot on 10 gestures, but then it sees a "thumbs down" (which it has never seen before), the old methods would force it to guess one of the 10 known gestures, likely getting it wrong.
The new method looks at the flashcard it built. It says, "This pattern doesn't match any of my 10 flashcards." It can confidently say, "I don't know what this is." This is crucial for real-world safety, like a self-driving car realizing it's seeing a strange new obstacle it wasn't trained on.

The Analogy: The Orchestra vs. The Soloist

Old Methods (Rate/Latency): Like a soloist playing a song. If the soloist misses a note or plays too slowly, the whole performance fails.
New Method (SNN-HDC): Like a massive orchestra playing a symphony. Even if a few instruments are slightly off or quiet, the melody is still clear to the audience. The "song" (the data) is represented by the collective harmony of hundreds of instruments, making it impossible to mess up and easy to recognize instantly.

The Bottom Line

The researchers at the University of Manchester have built a new "translator" for robot brains. It allows them to understand the world faster, with less battery power, and with the ability to admit when they are confused about something new. It's a big step toward making AI that is as efficient and adaptable as the human brain.

1. Problem Statement

Spiking Neural Networks (SNNs) are the third generation of neural networks, inspired by biological brains, offering potential for superior energy efficiency compared to Artificial Neural Networks (ANNs). However, their widespread adoption is hindered by:

Lower Accuracy: SNNs often underperform ANNs in learning tasks.
Inefficient Decoding: Current decoding methods have significant trade-offs:
- Rate Decoding: Relies on counting spikes over a time window. While accurate, it requires high spike counts (high energy) and high latency (waiting for the window to close).
- Latency Decoding: Relies on the timing of the first spike. It is faster and more energy-efficient but suffers from lower accuracy and poor noise robustness.
Lack of Distributed Representations: Most SNNs use one-hot encoding (one neuron per class), which limits expressiveness and robustness. Biological brains likely use distributed representations, which current SNN architectures do not fully exploit.
Inability to Detect Unknowns: Standard SNNs cannot easily identify classes they were not trained on.

The paper aims to develop a decoding method that achieves high accuracy, low latency, low energy consumption, and high noise robustness while enabling the detection of unknown classes.

2. Methodology

The authors propose a novel architecture called SNN-HDC, which combines Spiking Neural Networks with Hyperdimensional Computing (HDC).

Core Concept

Instead of outputting a single "winner-take-all" neuron (one-hot), the SNN-HDC outputs a hypervector (a high-dimensional binary vector).

Architecture: The SNN uses Leaky-Integrate-and-Fire (LIF) neurons. The output layer consists of $D$ neurons, where $D$ is the dimensionality of the hypervector (experimentally determined, e.g., 1024).
Encoding Mechanism:
- Each output neuron corresponds to one dimension in the hypervector.
- Dimensions are initialized to 0.
- If an output neuron fires a spike during the input sample, the corresponding dimension flips from 0 to 1.
- This creates a binary hypervector representing the temporal features observed.
Training & Loss:
- Target class hypervectors are generated randomly (50% probability for 0 or 1) to ensure they are pseudo-orthogonal.
- The loss function is Mean Squared Error (MSE) between the output hypervector and the target class hypervector.
- Crucially, the loss is calculated based on the presence of spikes (binary 0/1), not the count. If a target dimension is 1, multiple spikes do not increase loss; if 0, spikes increase loss.
Inference:
- Classification is performed by calculating the Hamming distance between the output hypervector and the stored class hypervectors.
- The class with the minimum Hamming distance is selected.
- This allows for event-driven, continuous classification (decisions can be made as the hypervector settles, without waiting for a fixed time window).

Experimental Setup

Datasets: DvsGesture (hand gestures) and SL-Animals-DVS (sign language).
Comparisons: The SNN-HDC was compared against:
- Rate-1 / Latency-1: Same depth as SNN-HDC but with standard one-hot output layers.
- Rate-2 / Latency-2: Deeper architectures (adding a hidden layer) with one-hot output layers.
Metrics: Classification accuracy, estimated energy consumption (based on Synaptic Operations - SOPs), classification latency, and spike counts.

3. Key Contributions

Novel Decoding Paradigm: Introduced a direct SNN-to-HDC mapping where the network is trained to output hypervectors, eliminating the need for a separate feature extraction stage or matrix multiplication for encoding (unlike previous works like SpikeHD or HyperSpike).
Energy Efficiency: By utilizing binary hypervectors and Hamming distance (addition-based logic) rather than matrix multiplication, the method is highly compatible with neuromorphic hardware.
Unknown Class Detection: Demonstrated that the SNN-HDC can identify unseen classes by measuring the dissimilarity (Hamming distance) of the output to known class hypervectors, a capability inherent to distributed representations but absent in one-hot SNNs.
Temporal Resolution: The model processes data at 1ms resolution without temporal compression, preserving fine-grained temporal information often lost in other SNN studies.

4. Results

The SNN-HDC was evaluated on DvsGesture and SL-Animals-DVS datasets.

Performance on DvsGesture

Accuracy: Achieved 96.59% accuracy. This is higher than Rate-1 (95.08%) and Latency models, and only 0.76% lower than the deepest Rate-2 model (97.35%).
Energy Consumption: Achieved the lowest energy usage.
- 1.24× to 3.67× reduction in estimated energy compared to Rate-1, Rate-2, and Latency models.
- Total energy: 2.54 mJ per sample (vs. 9.31 mJ for Rate-2).
Latency: Average latency of 162 ms. This is faster than Rate-1 and Latency-1, though slightly slower than the deeper Rate-2 model (133 ms).
Spike Count: Significantly reduced spike counts (1.74× to 3.36× fewer spikes) compared to other models.

Performance on SL-Animals-DVS

Accuracy: Achieved 74.13%, outperforming all other models including Rate-2 (72.74%) and Latency models.
Energy: 1.38× to 2.27× energy reduction compared to counterparts.
Latency: Lower latency than both Rate-1 and Rate-2 models.

Unknown Class Detection

When trained on 10 classes and tested on 11 (including an unseen "random" class), the SNN-HDC could identify 100% of the unknown samples when the similarity threshold ( $\delta$ ) was set to 0.1, while maintaining 94.17% accuracy on known classes.

5. Significance and Conclusion

This paper presents a compelling alternative to traditional rate and latency decoding in SNNs. The SNN-HDC model successfully bridges the gap between biological plausibility (distributed representations) and engineering efficiency.

Superior Trade-offs: It breaks the typical accuracy-energy-latency trade-off, offering high accuracy with significantly lower energy consumption and competitive latency.
Robustness: The use of high-dimensional vectors provides inherent noise robustness, allowing the model to perform well even on noisy neuromorphic data.
Scalability: The paper highlights that for large numbers of classes (exceeding ~2633), hyperdimensional encoding becomes more parameter-efficient than one-hot encoding.
Future Impact: The ability to detect unknown classes and process continuous event streams without resetting makes this approach highly suitable for real-world neuromorphic applications, such as autonomous robotics and edge AI, where energy and latency are critical constraints.

The authors conclude that combining SNNs with Hyperdimensional Computing is a viable path toward maturing neuromorphic algorithms, offering a robust, energy-efficient, and biologically inspired decoding strategy.